Right-Sizing UPS for the SMB Server Room

A typical SMB server room today draws 5 to 10 kW — two or three Proxmox hosts, a TrueNAS with a decent number of spindles, an OPNsense appliance, a 24-port switch with PoE and the backup server that hoovers everything up at night. When the power drops in that setup, seconds decide whether Monday morning brings a clean boot or a shredded ZFS pool. The uninterruptible power supply is no longer insurance — it is part of the operational baseline.

Even so, we regularly see two extremes in the field: either the 1500 VA tower UPS from the consumer electronics shop that keeps a single host alive for twelve minutes, or the 20 kVA unit from an old project that oversupplies two racks and whose batteries have not been swapped in four years. This article clears out both extremes and shows how to properly size a UPS for a typical SMB server room.

Measure the load profile — do not guess

The first mistake happens before the UPS is even ordered: estimated loads are added up from power supply labels. An 800 W PSU rarely draws more than 200 to 300 W under real load. Adding nameplate ratings gets you twice the capacity you need — and you pay for that three times over in acquisition, batteries and cooling.

Measure real active power at the rack, ideally over a week with a metered PDU or a clamp meter on the feed. A typical SMB setup shows values like this:

Component	Nameplate	Real active load
2x Proxmox host (24 cores)	2x 800 W	2x 220 W
TrueNAS with 12 HDDs	750 W	210 W
Backup server	500 W	90 W
OPNsense appliance	120 W	45 W
Switch 24 port + 8 PoE+	370 W	180 W
Misc — KVM, monitor	100 W	40 W
Total	3140 W	1005 W

For sizing, add a 25 to 30 percent reserve factor to the measured load for peaks (backup windows, RAID rebuilds, PoE peaks). So 1000 W real load becomes about 1300 W design load — not 3100 W.

On the UPS side, look at active power in watts, not apparent power in VA. Modern PSUs with active PFC run at power factor 0.95 to 0.99, but many UPS datasheets still assume 0.7, which massively distorts the comparison.

Runtime targets: 10 minutes for shutdown, 60 minutes to ride out

The runtime question is not “how long can the UPS carry the load” but “which scenario do I want to cover”. In our Proxmox consulting work we use two target values that have proven themselves in almost every SMB:

10 minutes for a graceful shutdown. That is the floor. In this window all VMs must be shut down in order, ZFS pools exported and hosts powered off cleanly. 10 minutes are enough for an orchestrated shutdown of a medium cluster with 30 to 50 VMs, provided the shutdown order is clearly defined.

60 minutes to ride out. Statistically, 60 minutes cover well over 90 percent of grid outages in Germany. Short flicker events, distribution grid switching, tripped breakers — these are usually resolved within 20 to 40 minutes. With 60 minutes of buffer you do not have to trigger an emergency shutdown at every hiccup.

For 1300 W design load with 60 minutes of runtime you need a UPS around 3000 VA with an external battery pack, or a 5 kVA class unit with internal extended batteries. The difference is less about the UPS price itself and more about the batteries — and those are the recurring cost item.

If you need very long runtimes (site without a diesel genset, rural grid connection), you should seriously consider a small diesel or gas generator once planned bridging exceeds roughly 30 minutes. The UPS then only has to bridge the generator start phase (typically 30 to 90 seconds).

Line-interactive vs online double-conversion

The second big design decision is topology. Short version:

Line-interactive (also VI, voltage independent): The load is normally connected directly to the grid, the UPS only corrects voltage variations via an autotransformer. On grid loss it switches to battery in typically 2 to 6 milliseconds. Transfer time is uncritical for every modern server PSU. Efficiency in normal operation: 96 to 98 percent.

Online double-conversion (VFI): The load is permanently fed from the UPS inverter, the grid only charges the batteries. There is no transfer, output voltage is always cleanly sinusoidal and frequency-stable. Efficiency depending on model and load 92 to 96 percent, slightly below that in ECO mode.

For a typical SMB server room with stable grid feed, line-interactive is usually sufficient and more economical. Online makes sense when:

• grid quality is demonstrably poor (frequent outages, voltage variation beyond +/- 10 percent, frequency drift from local generation),
• you integrate a generator and need a clean transfer buffer,
• sensitive medical or measurement equipment shares the circuit,
• the server room is operated under medical, industrial or higher-availability class requirements.

For 90 percent of our SMB projects we deploy line-interactive in rack format (2 to 3 U) plus an external battery pack. The price-performance ratio works out, and the efficiency advantage saves a four-digit euro amount in electricity over five years.

Battery lifetime and replacement cycle

The UPS itself lasts 10 to 15 years. The batteries do not. That is the second common mistake — the unit is installed, everything runs, nobody thinks about the batteries until the first real test comes and the UPS gives up after two minutes.

Lead-acid batteries in UPS use typically last 3 to 5 years, depending on ambient temperature and cycle count. Rule of thumb: for every 10 degrees Celsius above 20, lifetime halves. A server room running steadily at 24 degrees noticeably eats into battery life compared to a 20 degree room.

Practical approach:

• Calendar entry for battery swap after 4 years, regardless of apparent condition
• Semi-annual self-test via UPS web interface or NUT
• Annual full-load battery test (short discharge under real load), documented
• Keep ambient temperature under 22 degrees — that is cooling, not full climate control

For LiFePO4-based UPS (economically available at medium power ratings since around 2023) the picture shifts: lifetime 8 to 10 years, significantly higher cycle count, less temperature sensitivity. Acquisition is roughly 40 percent more expensive, but over the lifetime LiFePO4 is often already cheaper than the second lead-acid battery replacement after eight years.

NUT for auto-shutdown of Proxmox and TrueNAS

A UPS without shutdown automation is just an expensive delay. The glue between UPS and the connected systems is NUT — Network UPS Tools. NUT runs on Debian, Proxmox VE, TrueNAS SCALE and practically every Linux distribution. The UPS is connected via USB or SNMP, one master notifies all clients on the LAN.

A typical setup: the UPS is connected via USB to Proxmox host 1, which acts as NUT master. Proxmox host 2, TrueNAS and the backup server are NUT slaves and poll status over TCP.

Minimal /etc/nut/ups.conf on the master:

[serverroom]
    driver = usbhid-ups
    port = auto
    desc = "Eaton 9PX 3000 Rack"

/etc/nut/upsmon.conf on a slave (TrueNAS SCALE, also configurable via the UI):

MONITOR serverroom@192.168.10.11 1 monuser secret slave
SHUTDOWNCMD "/sbin/shutdown -h +0"
NOTIFYCMD /usr/sbin/upssched
FINALDELAY 5

More important than the configuration itself is the shutdown order. In our TrueNAS practice we define it like this:

1. On LOWBATT — when remaining runtime or battery voltage drops below threshold: all VMs on the Proxmox hosts shut down in order (4 minute timeout)
2. Then the Proxmox hosts themselves (2 minute timeout)
3. Finally TrueNAS and the backup server (2 minute timeout)
4. Only then does the master send the killpower command to the UPS

This order ensures storage systems are still alive when the VMs sync their disks. Shutting down TrueNAS first risks hung iSCSI or NFS sessions and, in the worst case, inconsistent VM disks.

Test the whole thing at least once per quarter — pull the UPS mains plug, run the stopwatch, document. Without a test the finest NUT configuration is just theory.

Rack integration, wiring and pitfalls

A few practical points that often slip through in the tender phase:

• Two PSUs, two feeds. Servers with redundant power supplies only actually add value when the A and B side hang on different UPS units, or at minimum on different output groups. Otherwise the redundancy is cosmetic.
• Plan for a maintenance bypass. Without an external maintenance bypass you have to shut down the entire load for a battery swap. A manual bypass at 200 to 500 euro saves you an after-hours maintenance window at every service call.
• Split output groups sensibly. Modern UPS units have switchable output groups. Put non-critical loads (monitors, test systems) on a group that is shed at 30 percent remaining runtime — that extends runtime for the systems that really matter.
• Include PoE in the calculation. A fully populated PoE+ switch quickly draws 300 to 500 W. If you only size for servers and forget IP telephony or access points, you are easily 20 percent off.

Conclusion

Right-sizing a UPS for the SMB server room is not rocket science, but it demands real measurements, clear runtime targets and a plan for orderly shutdown. Line-interactive is almost always the more economical choice for stable grids, 10 minutes of shutdown time plus 60 minutes of bridging is the proven target corridor, and batteries are a recurring maintenance item — not a one-off investment.

DATAZONE supports you with sizing, procurement and integration of your UPS infrastructure — from active-power measurement through NUT integration to regular battery and load testing. In our customer projects we have been operating mixed Proxmox, TrueNAS and OPNsense environments with Eaton, APC and Riello UPS units for years and know the pitfalls from practice. Get in touch — Contact DATAZONE.